Resumen de A connectionist architecture for hybrid training
We present a Connectionist architecture called Modified Hyperspherical Classifier (MHC) based on: 1) the work of Cooper [5] [6] and Batchelor [2] [3] [4] about Hyperspherical Classifiers. 2) the RCE paradigm as described by Scofield et al. [10] and 3) some new considerations derived from the search of an efficient model to perform heterogeneous pattern processing using supervised and/or unsupervised training algorithms depending on the nature of the problem, the availability of the correct output during the training process and the operational state of the network. We use the term Hybrid Trainig to define the use of a supervised or an unsupervised strategy to train the same network; this definition differs from the presented by Hertz et al. [7] as Hybrid Learning, wich refers to layered netwoks with different learning strategies for each layer (e. g. Counterpropagation).
Conclusions about the performance of the architecture are presented, based on the execution of tests using tasks related with Familiarty detection. Principal component analysis, Clustering, Prototyping, Feature mapping and Pattern transformation
